(D-154) Categorizing Data Acquisition System for Head Response Head Impact Data with Deep Learning

Introduction::

Mild traumatic brain injury (mTBI) or “concussive” injuries are a major societal issue and are associated with activities such as sports, motor vehicle crashes, and falls.  Quantifying head injury in high school tackle football is essential for determining a player’s injury risk per athletic exposure. However, the high prevalence of false positives in recorded head impact data means that data processing is critical to drawing accurate conclusions.

Materials and Methods:: In the current study, deep learning techniques were used to discriminate between true and false positive impact events using simulated football data. Using the Data Acquisition System for Head Response (DASHR) earpiece sensor, head impact data was recorded for four conditions: 1) high-g non-impacts, 2) valid impacts, 3) standing still, and 4) walking/jogging. Helmeted drop tests were performed with human surrogate heads to simulate the valid head impact events, while the high-g false positive events were simulated by pressing, flicking, and re-situating the DASHR in the ear. Walking/jogging and standing still were also recorded and were included in the model to encompass a broader range of activities. A 2-layer convolutional neural network (CNN) was trained and validated on 453 samples and evaluated on a test set of simulated data.

Results, Conclusions, and Discussions:: The model, a 2-layer convolutional neural network (CNN), was trained and validated on the 453 samples and achieved an accuracy of 84.6 ± 2.0% on a test set of simulated data.  Due in part to uneven sizes in the categorical training data, varying levels of sensitivity (77.8%, 90.1%, 40%, 87.5%) and precision (77.8%, 92.8%, 100%, 46.7%) were achieved for high-g non-impacts, valid impacts, standing still, and walking/jogging, respectively. Future work includes manually processing video footage from high school football practices and games to create a valid, robust training set to replace the simulated data. This data will then be augmented and used to train the current model’s architecture to create a tool for processing real-world head impact data from high school football players.

Acknowledgements (Optional): : Bass Connections at Duke University, Duke Institute for Brain Sciences and USA Lacrosse

References (Optional): :